A deep space debris anomaly detection method and system

Abstract

The application discloses a deep space debris anomaly detection method and system, belongs to the technical field of anomaly detection, acquires an original deep space exploration image, filters pure background pixels from the original deep space exploration image based on Mahalanobis distance measurement and spatial continuity verification, and constructs a training set; the pure background pixels after adding random Gaussian noise are used as noise-containing samples, input into a denoising network to obtain predicted noise, and a difference between the predicted noise and the Gaussian noise is used to construct a denoising loss; the network is iteratively optimized based on the denoising loss and a reconstruction loss to obtain a trained denoising network; a deep space exploration image to be detected is input into the trained denoising network to reconstruct a pure background image; Mahalanobis distance detection is performed on the reconstructed pure background image to determine an anomaly score of each pixel in the reconstructed pure background image; and when the anomaly score is higher than a preset threshold, it is determined that the pixel corresponds to a space debris. The denoising network constructed by the method can effectively improve the deep space debris detection precision.

Description

Technical Field

[0001] This invention relates to the field of anomaly detection technology, and more specifically to a method and system for detecting deep space debris anomalies. Background Technology

[0002] In deep space exploration, real-time anomaly detection of space debris is a core component of ensuring the safety of space missions. Space debris possesses complex physical characteristics, exhibiting extremely high speeds, small sizes, irregular shapes, and complex trajectories. Furthermore, the complex background environment of deep space makes debris easily obscured by background noise, significantly increasing the difficulty of detection.

[0003] Existing deep-space debris detection methods mainly rely on traditional image processing techniques, such as thresholding and morphological operations, or utilize two-dimensional convolutional neural networks. However, in practical engineering scenarios, traditional two-dimensional detection methods can only extract planar spatial features, resulting in a small receptive field. They cannot deeply explore the differences between debris and complex deep-space backgrounds in multi-dimensional feature space, leading to low detection accuracy and a high false negative rate for small targets. Furthermore, existing methods often directly use original images containing potential debris or strong noise for training, making the models highly susceptible to learning contamination distributions that include interference terms. This makes it difficult for the models to distinguish between real debris and background noise in practical applications. Summary of the Invention

[0004] To address the problems existing in the above-mentioned fields, this invention proposes a method and system for detecting deep space debris anomalies. The constructed denoising network can effectively remove potential debris interference in deep space images, ensuring that the training samples only contain a clean background distribution and avoiding model learning of contaminated distributions.

[0005] To address the aforementioned technical problems, this invention discloses a method for detecting deep space debris anomalies, comprising the following steps: We acquire raw deep space exploration images, and based on Mahalanobis distance metric and spatial continuity verification, we select clean background pixels from the raw deep space exploration images to construct a training set. A denoising network is constructed based on the 3D-UNet network architecture, including an encoder, a bottleneck layer, and a decoder. A multi-scale feature fusion module and a residual block are added sequentially between the input layer and the downsampling module of the original 3D-UNet network encoder, and an edge enhancement module is integrated in the residual block. Specifically, the multi-scale feature fusion module extracts multi-scale spatial-spectral features from the input data; the edge enhancement module enhances the edges of the multi-scale spatial-spectral features; the downsampling module uses a parallel-branch convolutional structure to extract the downsampling features of the enhanced features and the spatial dimension detail features, which are then fused to obtain the fused features; the bottleneck layer extracts the global semantic features of the fused features, which are then processed by the decoder to output predicted noise. The clean background pixels with randomly added Gaussian noise are used as noisy samples and input into the denoising network to obtain the predicted noise. The difference between the predicted noise and the Gaussian noise is used to construct the denoising loss. The background image is reconstructed based on the predicted noise and the difference between the reconstructed background image and the clean background pixels is used to construct the reconstruction loss. The denoising network is iteratively optimized through the denoising loss and the reconstruction loss to obtain the trained denoising network. The deep space exploration image to be detected is input into the trained denoising network, which outputs predicted noise. Based on the predicted noise and the deep space exploration image to be detected, the reconstructed clean background image is estimated. Mahalanobis distance is used to detect the reconstructed clean background image to determine the anomaly score of each pixel in the reconstructed clean background image. When the anomaly score is higher than a preset threshold, the pixel is determined to correspond to space debris.

[0006] Preferably, the step of acquiring multiple noisy deep space exploration images, and selecting clean background pixels from the noisy deep space exploration images based on Mahalanobis distance metric and spatial continuity verification to construct a training set specifically includes: Acquire multiple raw deep space exploration images; Determine the Mahalanobis distance of each pixel in each original deep space exploration image; based on a preset percentile threshold, select the pixel with the smallest Mahalanobis distance as multiple candidate background pixels; extract the spatial neighborhood window of each candidate background pixel and determine the proportion of candidate background pixels in that neighborhood; When the proportion is greater than the preset proportion threshold, the candidate background pixel is determined to be a clean background pixel, and a clean background mask is generated. Background image patches are extracted from each original deep space exploration image based on a clean background mask, and used as training samples.

[0007] Preferably, the multi-scale feature fusion module extracts multi-scale spatial-spectral features from the input data, specifically including: The multi-scale feature fusion module includes parallel convolutional layers and pooling layers; The convolutional layer uses parallel three-dimensional convolutional kernels of different scales to extract local features of pixels at different scales in noisy deep space images, and performs linear interpolation on the local features at different scales to obtain the corresponding spatial-spectral features of the input data. The pooling layer extracts global background features of pixels in noisy deep space images through global average pooling; Furthermore, a multi-scale attention mechanism is used to fuse spatial-spectral features and global background features to obtain multi-scale spatial-spectral features.

[0008] Preferably, the edge enhancement module enhances the edges of multi-scale spatial-spectral features, specifically including: The edge enhancement module employs a multi-directional edge detection convolution kernel, which learns the weights of multi-scale spatial-spectral features. By weighting and fusing multi-scale spatial-spectral features with edge features in various directions, enhanced features are obtained by weighting the multi-scale spatial-spectral features.

[0009] Preferably, the downsampling module uses a parallel branch convolutional structure to extract the downsampling features of the enhanced features and the spatial dimension detail features respectively, and then fuses them to obtain the fused features, specifically including: The downsampling module includes a convolutional structure with parallel branches and an attention mechanism layer, wherein the convolutional structure includes standard three-dimensional convolution and pseudo two-dimensional convolution; The downsampled features and spatial dimension detail features of the enhanced features are extracted by parallel extraction of standard 3D convolution and pseudo 2D convolution. By fusing downsampled features and spatial dimension detail features through an attention mechanism layer, fused features are obtained.

[0010] Preferably, after the bottleneck layer extracts the global semantic features of the fused features, the decoder processes the noise to output prediction noise, specifically including: The bottleneck layer is composed of multiple stacked residual blocks and integrates an efficient attention mechanism, wherein: The residual block performs high-order abstraction and nonlinear transformation on the fused features; The efficient attention mechanism establishes global associations of the transformed fused features, captures long-range dependencies of feature maps to model the complex statistical characteristics of the deep space background, and then extracts global semantic features of the fused features. The decoder includes an upsampling module, skip connections, and an output layer, wherein: The upsampling module uses trilinear interpolation to upsample the global semantic features, restores the feature map resolution, and obtains the upsampled features. The skip connection concatenates the shallow features corresponding to the encoder with the upsampled features in the channel dimension and performs size alignment to obtain fused detail features. The output layer processes the fused detail features through 3D convolution and outputs predicted noise.

[0011] Preferably, the step of using clean background pixels with randomly added Gaussian noise as noisy samples and inputting them into a denoising network to obtain predicted noise; constructing a denoising loss based on the difference between the predicted noise and the Gaussian noise; reconstructing the background image based on the predicted noise estimate; constructing a reconstruction loss based on the difference between the reconstructed background image and the clean background pixels; and iteratively optimizing the denoising network through the denoising loss and the reconstruction loss to obtain the trained denoising network, specifically includes: The denoising network is used as a parameterized denoiser for the diffusion model; Obtain the diffusion time step of the diffusion model A noise scheduling scheme is constructed, and the cumulative coefficient is determined using linear or cosine scheduling strategies. ,in, For noise dispatch coefficient, ; For clean background pixels in the training set, random sampling time steps and Gaussian noise Gaussian noise is added to the clean background pixel using the forward formula of the diffusion model to generate a noisy sample; The forward formula for the diffusion model is: ; in, For the clean background pixels in the training set, For the generated noisy samples; Noisy samples and their corresponding time steps Input the noise into the denoising network and obtain the predicted noise through forward propagation. ; The background image is then estimated using the predicted noise and accumulation coefficients obtained from forward propagation. ; The composite loss function is constructed as follows: ; in, For noise reduction loss; For reconstruction loss; Based on the composite loss function, the network parameters of the denoising network are updated through backpropagation until convergence, resulting in the trained denoising network.

[0012] Preferably, the step of inputting the deep space exploration image to be detected into the trained denoising network and outputting predicted noise; and estimating and reconstructing a clean background image based on the predicted noise and the deep space exploration image to be detected, specifically includes: The trained denoising network is used as the denoiser for the diffusion model; The deep space exploration image to be detected is converted into tensor format and Gaussian noise is randomly added to generate noisy samples, which serve as the initial state for the diffusion model. ; Based on the forward propagation of the diffusion model, the diffusion model performs iterative inference on the deep space exploration image to be detected, and performs multiple iterations of denoising to gradually remove noise: For each iteration k =1 to K, the current noisy state Input the trained denoising network to predict the noise in the current state. ; Backpropagation based on a diffusion model reconstructs the background of the denoised current state using noise estimation from the predicted current state. ; And based on the background of the reconstructed and denoised current state Update the input state of the diffusion model: ; After K iterations, the final output is a reconstructed clean background image. ,in H, W Image spatial resolution, C This refers to the number of bands.

[0013] Preferably, the step of performing Mahalanobis distance detection on the reconstructed clean background image to determine the anomaly score of each pixel in the reconstructed clean background image specifically includes: Mahalanobis distance detection is performed on the reconstructed image to obtain the anomaly score of each pixel in the reconstructed clean background image: ; in, Represents the first element in the reconstructed clean background image. i 1 pixel; This represents the mean value of each pixel in the reconstructed clean background image; T This indicates that the transpose condition is met; Represents the covariance matrix The inverse matrix.

[0014] Preferably, it also includes a deep space debris anomaly detection system, comprising: The data acquisition module is used to acquire raw deep space exploration images. Based on Mahalanobis distance metric and spatial continuity verification, it filters clean background pixels from the raw deep space exploration images to construct a training set. A denoising network construction module is used to construct a denoising network. The denoising network is based on the 3D-UNet network architecture and includes an encoder, a bottleneck layer, and a decoder. A multi-scale feature fusion module and a residual block are added sequentially between the input layer and downsampling module of the original 3D-UNet encoder, and an edge enhancement module is integrated into the residual block. Specifically, the multi-scale feature fusion module extracts multi-scale spatial-spectral features from the input data; the edge enhancement module enhances the edges of the multi-scale spatial-spectral features; the downsampling module uses a parallel-branch convolutional structure to extract the downsampling features of the enhanced features and the spatial dimension detail features, which are then fused to obtain the fused features; the bottleneck layer extracts the global semantic features of the fused features, which are then processed by the decoder to output predicted noise. The denoising network training module is used to take clean background pixels with randomly added Gaussian noise as noisy samples and input them into the denoising network to obtain predicted noise; construct a denoising loss based on the difference between the predicted noise and the Gaussian noise; reconstruct the background image based on the predicted noise estimate and construct a reconstruction loss based on the difference between the reconstructed background image and the clean background pixels; and iteratively optimize the denoising network through the denoising loss and the reconstruction loss to obtain the trained denoising network. The deep space debris anomaly detection module is used to input the deep space exploration image to be detected into the trained denoising network and output predicted noise; based on the predicted noise and the deep space exploration image to be detected, it estimates and reconstructs a clean background image; it performs Mahalanobis distance detection on the reconstructed clean background image to determine the anomaly score of each pixel in the reconstructed clean background image; when the anomaly score is higher than a preset threshold, the pixel is determined to be a space debris.

[0015] Compared with the prior art, the present invention has the following beneficial effects: The proposed deep-space debris anomaly detection method effectively eliminates potential debris interference and isolated noise points in the original deep-space image by using Mahalanobis distance metric and spatial continuity verification to filter clean background pixels. This ensures the purity of the training samples, allowing the denoising network to learn only the statistical distribution patterns of the clean deep-space background during training. This avoids the model mistaking debris features for background features and solves the pollution distribution problem at its source. The constructed denoising network extracts features at different scales in parallel through a multi-scale feature fusion module, expanding the network's receptive field. The edge enhancement module uses multi-directional edge detection convolutional kernels, significantly enhancing the sensitivity to weak debris edge features. During network training, the constructed denoising loss and reconstruction loss constrain both denoising capability and reconstruction fidelity. The denoising loss provides coarse-grained backpropagation gradients, guiding the network to quickly locate and remove noise distributions; the reconstruction loss provides fine-grained pixel-level supervision, ensuring the denoised background remains undistorted. These two losses complement each other during training iterations, significantly accelerating the convergence speed of the denoising network and greatly improving the signal-to-noise ratio and visual fidelity of the final reconstructed clean background image, making subsequent Mahalanobis distance anomaly score calculations more accurate. The trained denoising network denoises the deep space exploration image to be detected, outputting predicted noise; based on the predicted noise and the deep space exploration image to be detected, a clean background image is estimated and reconstructed. This process accurately models complex deep space background distributions, significantly enhancing the feature discrimination between the background and debris targets. Attached Figure Description

[0016] Figure 1 This is an overall flowchart of the deep space debris anomaly detection method proposed in this invention; Figure 2 This is a schematic diagram of the process for selecting clean deep space background samples provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the overall architecture of the denoising network constructed in this invention; Figure 4 This is a schematic diagram of the multi-scale spatial semantic feature fusion and edge enhancement module in the denoising network provided in this embodiment of the invention; Figure 5 The original deep space test image provided in the embodiments of the present invention; Figure 6 The ground truth label map of the method provided in this embodiment of the invention and the comparison of space debris detection performance with other methods are shown in the figure. Figure 7 The ground truth label map of the method provided in this embodiment of the invention and the comparison of the space debris detection performance of other methods on the Debris-2 dataset are shown in the figure. Figure 8 Comparison of ROC curves of various methods provided in the embodiments of the present invention on the Debris-1 dataset; Figure 9 Comparison of ROC curves of various methods provided in the embodiments of the present invention on the Debris-2 dataset. Detailed Implementation

[0017] The following will refer to the appendices in the embodiments of the present invention. Figures 1-9 The technical solutions in the embodiments of the present invention will be clearly and completely described. It should be understood that the terminology used in the present invention is only for describing particular implementation methods and is not intended to limit the present invention.

[0018] Example like Figure 1 As shown, this invention proposes a method for detecting deep space debris anomalies, comprising the following steps: S1: Obtain the original deep space exploration images, and based on Mahalanobis distance metric and spatial continuity verification, select clean background pixels from the original deep space exploration images to construct a training set; S2: Construct a denoising network; The denoising network is based on the 3D-UNet network architecture, including an encoder, a bottleneck layer and a decoder. A multi-scale feature fusion module and a residual block are added sequentially between the input layer of the encoder and the downsampling module of the original 3D-UNet network, and an edge enhancement module is integrated in the residual block. The multi-scale feature fusion module extracts multi-scale spatial-spectral features from the input data; the edge enhancement module enhances the edges of the multi-scale spatial-spectral features; the downsampling module uses a parallel branch convolutional structure to extract the downsampling features of the enhancement features and the detailed features of the spatial dimension, and then fuses them to obtain the fused features; the bottleneck layer extracts the global semantic features of the fused features, and then the decoder processes the output prediction noise. S3: Take the clean background pixels after randomly adding Gaussian noise as noisy samples and input them into the denoising network to obtain the predicted noise; construct the denoising loss based on the difference between the predicted noise and the Gaussian noise; reconstruct the background image based on the predicted noise estimate and construct the reconstruction loss based on the difference between the reconstructed background image and the clean background pixels; iteratively optimize the denoising network through the denoising loss and the reconstruction loss to obtain the trained denoising network. S4: Input the deep space exploration image to be detected into the trained denoising network and output the predicted noise; estimate and reconstruct the clean background image based on the predicted noise and the deep space exploration image to be detected; perform Mahalanobis distance detection on the reconstructed clean background image to determine the anomaly score of each pixel in the reconstructed clean background image; when the anomaly score is higher than the preset threshold, the pixel is determined to correspond to space debris.

[0019] The embodiments of this invention use a deep space debris dataset. This dataset originates from the Debris Detection challenge on the AIcrowd international competition platform, and the acquisition of this data was legal and compliant.

[0020] To meet the requirements of three-dimensional spatial-spectral feature extraction of the method proposed in this invention, the LabelMe tool is used to perform fine pixel-level annotation on the fragment targets, and the annotation data is converted into .mat format using a Python script.

[0021] S11: Mahalanobis distance metric like Figure 2 As shown, the original deep space exploration image was obtained. In this embodiment, ,in H , W Image spatial resolution, C The number of bands is used as input to calculate the original deep space exploration image using Mahalanobis distance detection. X Mahalanobis distance of each pixel in This is used to measure the degree of deviation from background statistical characteristics. ; in, Raw deep space exploration images X The first in i Multi-band feature vector of 1 pixel ; μ It is the global mean vector; Let be the covariance matrix.

[0022] Based on a preset percentile threshold P% (95% in this embodiment), the pixels with the smallest Mahalanobis distance (the first (100-P)%) are selected as multiple candidate background pixels, and a binary candidate background pixel mask is generated. .

[0023] S12: Verification of Spatial Continuity For each candidate background pixel Extract the spatial neighborhood window of each candidate background pixel. In this embodiment, a 3×3 matrix is ​​used to extract the 8-neighborhood of the current candidate background pixel, count the number of candidate background pixels in the corresponding neighborhood, and calculate the proportion of candidate background pixels in the neighborhood. When the proportion is greater than the preset proportion threshold When the value is 0.5, the candidate background pixel is determined to belong to a continuous background region, marked as a clean background pixel, and a clean background mask is generated. This step can effectively eliminate isolated noise points or potentially tiny fragment pixels.

[0024] Based on a clean background mask, a set of clean background image patches is extracted from the original deep space exploration image X. Each image block is sized as follows: In this embodiment, a 3×3×3 sample is used as the training sample to ensure that the training data only contains the statistical characteristics of the deep space background and excludes interference from space debris.

[0025] S13: Data Preprocessing Normalize the clean background image patch, mapping the pixel values ​​to... Interval: ; Convert two-dimensional space-band data into a three-dimensional tensor format. ,in B For batch size, D To construct the depth dimension, this embodiment takes... B =1, D =6; The original dimension is reduced by trilinear interpolation. Expand to Specifically, this embodiment will use the original dimension Expand to Generate standard input tensor And as a background sample, this embodiment takes , as training samples, and construct the training set.

[0026] Denoising Network Construction like Figure 3 As shown, the denoising network constructed in this invention is based on the 3D-UNet network architecture, including an encoder, a bottleneck layer, and a decoder. A multi-scale feature fusion module and a residual block are added sequentially between the input layer and the downsampling module of the original 3D-UNet encoder. An edge enhancement module is integrated into the residual block; in this embodiment, the enhancement strength is set to 0.5. The constructed denoising network is used as a denoiser for a diffusion model.

[0027] This embodiment takes the base number of channels. ch =32, channel multiplier ch_mult =[1,2], a total of 2 downsampling stages, residual block repetition count 2, dropout rate set to 0.5.

[0028] S21: The encoder includes an input layer, a multi-scale feature fusion module, a downsampling module, and a residual block.

[0029] The input layer receives the image after noise addition. Image with added noise Specifically, this refers to the standard input tensor. Images generated from noisy samples after randomly adding sampled Gaussian noise; The multi-scale feature fusion module includes parallel convolutional layers and pooling layers, wherein: like Figure 4 As shown, the convolutional layer uses parallel 3D convolutional kernels of three scales: 1×1×1, 2×2×2, and 3×3×3 to extract the noisy image. Local features at different scales are obtained, and linear interpolation is performed on these local features to obtain the corresponding spatial-spectral features of the input data. Finally, noisy deep-space image pixels are extracted in the pooling layer using global average pooling. The global background features are obtained; and the spatial-spectral features and global background features are fused through a multi-scale attention mechanism to obtain multi-scale spatial-spectral features.

[0030] The residual block includes a time-step embedding module, a small target enhancement module, an edge enhancement module, and a gated linear unit attention mechanism, among which: The edge enhancement module employs multi-directional (spatial x / y direction, depth z direction, and diagonal direction) edge detection convolutional kernels. It fuses multi-scale spatial-spectral features in each direction using learnable weights, weighting these features to obtain enhanced features and improve sensitivity to weak debris edges. The temporal embedding module encodes the diffusion time step into a continuous vector representation and uses sinusoidal positional encoding to generate temporal features, guiding the denoising network to adapt to different noise levels.

[0031] The downsampling module includes convolutional structures with parallel branches and attention mechanism layers. The convolutional structures include standard 3D convolution (stride=2) and pseudo 2D convolution.

[0032] By using standard 3D convolution and pseudo 2D convolution in parallel, downsampled features and spatial dimension detail features are extracted to enhance the features.

[0033] At the attention mechanism layer, an attention mechanism guided by a temporal embedding module fuses downsampled features and spatial detail features, with the temporal embedding dimension set to... t dim =128, using sinusoidal position encoding to obtain fused features, preserving details while expanding the receptive field.

[0034] S22: The bottleneck layer is located at the center of the 3D-UNet network. It consists of two stacked residual blocks and integrates an efficient attention mechanism, in which: The residual blocks perform high-order abstraction and nonlinear transformation on the fused features; the efficient attention mechanism establishes global associations of the transformed fused features. The network can capture long-range dependencies of the feature maps across the entire feature map to model the complex statistical properties of the deep space background, and then extract global semantic features of the fused features.

[0035] S23: The decoder includes an upsampling module, skip connections, and an output layer, where: The upsampling module uses trilinear interpolation to upsample the global semantic features, restoring the feature map resolution and obtaining the upsampled features. Skip connections concatenate the shallow features corresponding to the encoder with the upsampled features along the channel dimension and align their sizes to ensure spatial dimension matching, preserving shallow detail information and obtaining fused detail features. The output layer processes the fused detail features through a 3×3×3 three-dimensional convolution, outputting prediction noise. , where θ is the network parameter of the denoising network.

[0036] S24: Use the constructed denoising network as the denoiser of the diffusion model to create the noisy image corresponding to the current state, i.e., the noisy sample. and time step t To predict noise The nonlinear mapping noise prediction function.

[0037] Network training S31: Using the denoising network constructed in step S2, build a diffusion model training framework, specifically including defining the diffusion time step. (In this embodiment, the value is 100000). A noise scheduling scheme is constructed, and the noise scheduling coefficient is calculated using a linear or cosine scheduling strategy. , from The cumulative coefficient is then calculated by linearly increasing the coefficient to 0.02. Noise Dispatch Coefficient The variance of the noise added to the input data at the t-th time step of the diffusion process is a predetermined noise variance coefficient.

[0038] For the clean background pixels selected in step S1, random sampling time steps are performed. and Gaussian noise Gaussian noise was added to the clean background pixel using the forward formula of the diffusion model to generate a noisy sample. for: ; Noisy samples and time embedding The denoising network constructed in step S2 is input, and the predicted noise is obtained through forward propagation. ; Based on the predicted noise, the cumulative coefficient during the forward propagation process is utilized. Perform the inverse operation from the noisy samples Subtracting the predicted noise yields a reconstructed clean background image. : ; S32: Construct a composite loss function and optimize the network parameters of the denoising network. : ; in, The denoising loss is used to constrain the denoising network to accurately estimate the Gaussian noise added during forward propagation. The reconstruction loss is used to constrain the reconstructed clean background image estimated by the denoising network to be consistent with the real clean background pixels. In this embodiment... Take 0.3.

[0039] The network parameters θ of the denoising network are updated by backpropagation algorithm, and iterative optimization is performed until convergence to obtain the trained denoising network, which can accurately predict the noise at each time step in the diffusion process, thereby achieving accurate modeling of the deep space background distribution.

[0040] Anomaly detection S41: The trained denoising network is used as a parameterized denoiser for the diffusion model. Iterative inference is performed on the deep space test image to be detected, and K iterations of denoising are performed (in this embodiment, K=100, batch size is 1, AdamW optimizer is used, and initial learning rate lr=1e-5) to gradually remove noise and reconstruct a clean background image.

[0041] Deep space exploration images to be detected (Size 512×512×3) is converted to the input tensor format [1,3,6,512,512] and noise is added as the initial state of the diffusion model. ; Based on the forward propagation of the diffusion model, the diffusion model performs iterative inference on the deep space exploration image to be detected, performing K=50 iterations of denoising to gradually remove noise: For each iteration k =1 to K, the current noisy state Input the trained denoising network to predict the noise in the current state. ; Backpropagation based on a diffusion model reconstructs the background of the denoised current state using noise estimation from the predicted current state. ; And based on the background of the reconstructed and denoised current state Update the input state of the diffusion model: ; After K iterations, the final output is a reconstructed clean background image. H=512, W=512, C=3.

[0042] Since the denoising network trained through step S4 has only learned the distribution of the clean deep space background, space debris, as an out-of-distribution object, cannot be accurately reconstructed by the trained denoising network, thus forming a significant residual in the reconstruction result.

[0043] S42: Perform Mahalanobis distance detection on the reconstructed clean background image to obtain the anomaly score of each pixel in the reconstructed clean background image: ; in, Represents the first element in the reconstructed clean background image. i 1 pixel; This represents the mean value of each pixel in the reconstructed clean background image; T This indicates that the transpose condition is met; Represents the covariance matrix The inverse matrix.

[0044] By setting a detection threshold or calculating the ROC curve and AUC value, when the abnormal score of each pixel in the reconstructed clean background image is greater than the preset threshold, the pixel is identified as a spatial fragment, thereby realizing the detection of abnormal targets in complex deep space backgrounds.

[0045] like Figure 5 Figure (a) shows the original deep space exploration image of the Debris-1 dataset provided in this embodiment, and Figure (b) shows the original deep space exploration image of the Debris-2 dataset provided in this embodiment.

[0046] like Figure 6Figure (a) shows the ground truth label of the Debris-1 dataset in this embodiment. Figures (b)-(f) show the results of deep space debris detection using different methods, where Figure (b) is the RX method, Figure (c) is the GMM method, Figure (d) is the U-Net method, Figure (e) is the BSDM method, and Figure (f) is the method proposed in this invention. Figures (b), (c), and (d) show many small bright spots in the detected background, which are false alarms. Figure (c) shows the most obvious false alarms. Figures (b) and (d) may appear less obvious due to being converted to black and white, and their backgrounds are not as clean as those detected in Figure (f). Therefore, their suppression of false alarms is not as good as the method proposed in this invention. Furthermore, the detection effect in Figure (d) shows a ring of false alarms around the target, indicating significant background influence and insufficient target detection accuracy. The detection effect in Figure (e) clearly shows that the target is not as bright as that in Figure (f), resulting in low distinction between the target and the background and insufficient target feature prominence.

[0047] like Figure 7 Figure (a) shows the ground truth label of the Debris-2 dataset in this embodiment. Figures (b)-(f) show the results of deep space debris detection using different methods, where Figure (b) is the RX method, Figure (c) is the GMM method, Figure (d) is the U-Net method, Figure (e) is the BSDM method, and Figure (f) is the method proposed in this invention.

[0048] This embodiment uses the ROC (Receiver Operating Characteristic) curve and AUC (Area Under Curve) value as detection indicators, wherein: ROC (FPR, TPR) describes the relationship between the false positive rate (FPR) and the true positive rate (TPR) at different thresholds. The closer the curve is to the upper left corner of the coordinate axis, the higher the detection rate.

[0049] Under certain thresholds, the formulas for calculating FPR and TPR are as follows: ; ; Where TP (True Positive) and TN (True Negative) represent the number of detected true anomalies and background pixels, respectively. FP (False Positive) and FT (False Negative) represent the number of pixels falsely detected as anomalies and background pixels, respectively. The AUC value is the area under the ROC curve.

[0050] like Figure 8The figure shown is a comparison of the ROC curves of the proposed method in this embodiment with several other methods on the Debris-1 dataset.

[0051] Figure 9 This is a comparison of the ROC curves of the method proposed in this embodiment with several other methods on the Debris-2 dataset.

[0052] Analyze separately Figure 8 and Figure 9 As can be seen from the curve trends of different methods, the detection rate of the ROC curve of the method proposed in this invention is closer to 1 than that of other methods, and the false alarm rate is closer to 0 than that of other methods, indicating that the detection performance of the method proposed in this invention is better.

[0053] Table 1 shows the AUC values ​​of different datasets under RX and the method proposed in this invention, and Table 2 shows the FPR values ​​of different datasets under RX and the method proposed in this invention.

[0054] Table 1. AUC values ​​for different datasets under RX and the method proposed in this invention. Table 2. FPR values ​​for different datasets under RX and the method of this invention. As shown in Tables 1 and 2, the deep space debris anomaly detection method proposed in this invention exhibits excellent detection performance on both datasets, with an average detection rate (AUC) of 0.9986 and an average false alarm rate (FPR) as low as 0.0006. Compared with the comparative methods, this invention improves the detection accuracy by 0.0395, 0.0349, 0.0001, and 0.0217 compared to RX, GMM, U-Net, and BSDM, respectively, indicating that this invention has a stronger ability to distinguish complex backgrounds. In terms of false alarm rate control, this invention reduces the false alarm rate by 0.004, 0.0683, 0.0146, and 0.0024 compared to the above four methods, respectively, fully verifying the technical advantages of this invention in effectively suppressing background noise through pure background screening and three-dimensional diffusion reconstruction mechanism.

[0055] In summary, this invention maintains a high detection rate while significantly reducing the false alarm rate, and exhibits stable performance on multiple datasets. The variances of AUC and FPR are both less than 0.001, demonstrating good robustness and generalization ability, and can adapt to deep space exploration scenarios of varying complexity.

[0056] This invention also proposes a deep space debris anomaly detection system, comprising: The data acquisition module is used to acquire raw deep space exploration images. Based on Mahalanobis distance metric and spatial continuity verification, it filters clean background pixels from the raw deep space exploration images to construct a training set. A denoising network construction module is used to build the denoising network. The denoising network is based on the 3D-UNet network architecture, including an encoder, a bottleneck layer, and a decoder. A multi-scale feature fusion module and a residual block are added sequentially between the input layer and the downsampling module of the original 3D-UNet network encoder, and an edge enhancement module is integrated in the residual block. Specifically, the multi-scale feature fusion module extracts multi-scale spatial-spectral features from the input data; the edge enhancement module enhances the edges of the multi-scale spatial-spectral features; the downsampling module uses a parallel branch convolutional structure to extract the downsampling features of the enhanced features and the spatial dimension detail features, and then fuses them to obtain the fused features; the bottleneck layer extracts the global semantic features of the fused features, and then the decoder processes the output to produce predicted noise. The denoising network training module is used to take clean background pixels with randomly added Gaussian noise as noisy samples, input them into the denoising network to obtain predicted noise, construct the denoising loss based on the difference between the predicted noise and the Gaussian noise, and reconstruct the background image based on the predicted noise, construct the reconstruction loss based on the difference between the reconstructed background image and the clean background pixels. The denoising network is iteratively optimized through the denoising loss and the reconstruction loss to obtain the trained denoising network. The deep space debris anomaly detection module is used to input the deep space exploration image to be detected into the trained denoising network and output predicted noise; based on the predicted noise and the deep space exploration image to be detected, it estimates and reconstructs a clean background image; it performs Mahalanobis distance detection on the reconstructed clean background image to determine the anomaly score of each pixel in the reconstructed clean background image; when the anomaly score is higher than a preset threshold, the pixel is determined to be a space debris.

[0057] In summary, the deep space debris anomaly detection method proposed in this invention filters clean background pixels based on Mahalanobis distance metric and spatial continuity verification, extracts image patches to construct a training set, and performs preprocessing. A denoising network for a diffusion model is constructed to extract the three-dimensional spatial-spectral features of deep space data. This constructed denoising network is used as a parameterized denoiser for the diffusion model, forming a diffusion model training framework. Gradient information is generated through forward denoising and backward denoising processes, and network parameters are optimized via backpropagation to learn the background distribution. The deep space exploration image to be detected is input into the trained denoising network to reconstruct the background. Mahalanobis distance is then used to detect the reconstructed image to obtain an anomaly score, thus achieving space debris anomaly detection. This invention aims to accurately model complex deep space background distributions, significantly enhance the feature differentiation between background and debris targets, effectively improve the accuracy of deep space debris detection, and reduce the false alarm rate.

[0058] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0059] Furthermore, unless otherwise stated, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. All references to this specification are incorporated by way of citation to disclose and describe methods relating to those references. In the event of any conflict with any incorporated reference, the content of this specification shall prevail.

Claims

1. A method for detecting deep space debris anomalies, characterized in that, Includes the following steps: We acquire raw deep space exploration images, and based on Mahalanobis distance metric and spatial continuity verification, we select clean background pixels from the raw deep space exploration images to construct a training set. A denoising network is constructed; the denoising network is based on the 3D-UNet network architecture, including an encoder, a bottleneck layer and a decoder. A multi-scale feature fusion module and a residual block are added sequentially between the input layer and the downsampling module of the encoder of the original 3D-UNet network, and an edge enhancement module is integrated in the residual block; wherein, the multi-scale feature fusion module extracts multi-scale spatial-spectral features of the input data; The edge enhancement module enhances the edges of multi-scale spatial-spectral features; the downsampling module uses a parallel branch convolutional structure to extract the downsampling features of the enhanced features and the spatial dimension detail features respectively, and then fuses them to obtain the fused features; the bottleneck layer extracts the global semantic features of the fused features and then processes them through the decoder to output prediction noise; The clean background pixels with randomly added Gaussian noise are used as noisy samples and input into the denoising network to obtain predicted noise. The difference between the predicted noise and the Gaussian noise is used to construct the denoising loss. The background image is reconstructed based on the predicted noise and the difference between the reconstructed background image and the clean background pixels is used to construct the reconstruction loss. The denoising network is iteratively optimized through the denoising loss and the reconstruction loss to obtain the trained denoising network. The deep space exploration image to be detected is input into the trained denoising network, which outputs predicted noise. Based on the predicted noise and the deep space exploration image to be detected, the reconstructed clean background image is estimated. Mahalanobis distance is used to detect the reconstructed clean background image to determine the anomaly score of each pixel in the reconstructed clean background image. When the anomaly score is higher than a preset threshold, the pixel is determined to correspond to space debris.

2. The deep space debris anomaly detection method according to claim 1, characterized in that, The process of acquiring multiple noisy deep space exploration images, and then selecting clean background pixels from the noisy deep space exploration images based on Mahalanobis distance metric and spatial continuity verification to construct a training set, specifically includes: Acquire multiple raw deep space exploration images; Determine the Mahalanobis distance of each pixel in each original deep space exploration image; based on a preset percentile threshold, select the pixel with the smallest Mahalanobis distance as multiple candidate background pixels; extract the spatial neighborhood window of each candidate background pixel and determine the proportion of candidate background pixels in that neighborhood; When the proportion is greater than the preset proportion threshold, the candidate background pixel is determined to be a clean background pixel, and a clean background mask is generated. Background image patches are extracted from each original deep space exploration image based on a clean background mask, and used as training samples.

3. The deep space debris anomaly detection method according to claim 1, characterized in that, The multi-scale feature fusion module extracts multi-scale spatial-spectral features from the input data, specifically including: The multi-scale feature fusion module includes parallel convolutional layers and pooling layers; The convolutional layer uses parallel three-dimensional convolutional kernels of different scales to extract local features of pixels at different scales in noisy deep space images, and performs linear interpolation on the local features at different scales to obtain the corresponding spatial-spectral features of the input data. The pooling layer extracts global background features of pixels in noisy deep space images through global average pooling; Furthermore, a multi-scale attention mechanism is used to fuse spatial-spectral features and global background features to obtain multi-scale spatial-spectral features.

4. The deep space debris anomaly detection method according to claim 1, characterized in that, The edge enhancement module enhances the edges of multi-scale spatial-spectral features, specifically including: The edge enhancement module employs a multi-directional edge detection convolution kernel, which learns the weights of multi-scale spatial-spectral features. By weighting and fusing multi-scale spatial-spectral features with edge features in various directions, enhanced features are obtained by weighting the multi-scale spatial-spectral features.

5. The deep space debris anomaly detection method according to claim 1, characterized in that, The downsampling module uses a parallel branch convolutional structure to extract downsampling features of the enhancement features and spatial dimension detail features respectively, and then fuses them to obtain fused features, specifically including: The downsampling module includes a convolutional structure with parallel branches and an attention mechanism layer, wherein the convolutional structure includes standard three-dimensional convolution and pseudo-two-dimensional convolution; The downsampled features and spatial dimension detail features of the enhanced features are extracted by parallel extraction of standard 3D convolution and pseudo 2D convolution. By fusing downsampled features and spatial dimension detail features through an attention mechanism layer, fused features are obtained.

6. The deep space debris anomaly detection method according to claim 1, characterized in that, After the bottleneck layer extracts the global semantic features of the fused features, the decoder processes the noise to output prediction noise, specifically including: The bottleneck layer is composed of multiple stacked residual blocks and integrates an efficient attention mechanism, wherein: The residual block performs high-order abstraction and nonlinear transformation on the fused features; The efficient attention mechanism establishes global associations of the transformed fused features, captures long-range dependencies of feature maps to model the complex statistical characteristics of the deep space background, and then extracts global semantic features of the fused features. The decoder includes an upsampling module, skip connections, and an output layer, wherein: The upsampling module uses trilinear interpolation to upsample the global semantic features, restores the feature map resolution, and obtains the upsampled features. The skip connection concatenates the shallow features corresponding to the encoder with the upsampled features in the channel dimension and performs size alignment to obtain fused detail features. The output layer processes the fused detail features through 3D convolution and outputs predicted noise.

7. The deep space debris anomaly detection method according to claim 1, characterized in that, The process involves using clean background pixels with randomly added Gaussian noise as noisy samples, inputting them into a denoising network to obtain predicted noise; constructing a denoising loss based on the difference between the predicted noise and the Gaussian noise; reconstructing the background image based on the predicted noise estimate, and constructing a reconstruction loss based on the difference between the reconstructed background image and the clean background pixels; iteratively optimizing the denoising network using the denoising loss and the reconstruction loss to obtain the trained denoising network, specifically including: The denoising network is used as a parameterized denoiser for the diffusion model; Obtain the diffusion time step of the diffusion model A noise scheduling scheme is constructed, and the cumulative coefficient is determined using linear or cosine scheduling strategies. ,in, For noise dispatch coefficient, ; For clean background pixels in the training set, random sampling time steps and Gaussian noise Gaussian noise is added to the clean background pixel using the forward formula of the diffusion model to generate a noisy sample; The forward formula for the diffusion model is: ； in, For the clean background pixels in the training set, For the generated noisy samples; Noisy samples and their corresponding time steps Input the noise into the denoising network and obtain the predicted noise through forward propagation. ; The background image is then estimated using the predicted noise and accumulation coefficients obtained from forward propagation. ； The composite loss function is constructed as follows: ； in, For noise reduction loss; For reconstruction loss; Based on the composite loss function, the network parameters of the denoising network are updated through backpropagation until convergence, resulting in the trained denoising network.

8. The deep space debris anomaly detection method according to claim 1, characterized in that, The process involves inputting the deep space exploration image to be detected into a trained denoising network, which outputs predicted noise; and estimating and reconstructing a clean background image based on the predicted noise and the deep space exploration image to be detected. Specifically, this includes: The trained denoising network is used as the denoiser for the diffusion model; The deep space exploration image to be detected is converted into tensor format and Gaussian noise is randomly added to generate noisy samples, which serve as the initial state for the diffusion model. ; Based on the forward propagation of the diffusion model, the diffusion model performs iterative inference on the deep space exploration image to be detected, and performs multiple iterations of denoising to gradually remove noise: For each iteration k =1 to K, the current noisy state Input the trained denoising network to predict the noise in the current state. ; Backpropagation based on a diffusion model reconstructs the background of the denoised current state using noise estimation from the predicted current state. ; And based on the background of the reconstructed and denoised current state Update the input state of the diffusion model: ; After K iterations, the final output is a reconstructed clean background image. ,in H, W Image spatial resolution, C This refers to the number of bands.

9. The deep space debris anomaly detection method according to claim 1, characterized in that, The step of performing Mahalanobis distance detection on the reconstructed clean background image to determine the anomaly score of each pixel in the reconstructed clean background image specifically includes: Mahalanobis distance detection is performed on the reconstructed image to obtain the anomaly score of each pixel in the reconstructed clean background image: ； in, Represents the first element in the reconstructed clean background image. i 1 pixel; This represents the mean value of each pixel in the reconstructed clean background image; T This indicates that the transpose condition is met; Represents the covariance matrix The inverse matrix.

10. A deep space debris anomaly detection system, characterized in that, include: The data acquisition module is used to acquire raw deep space exploration images. Based on Mahalanobis distance metric and spatial continuity verification, it filters clean background pixels from the raw deep space exploration images to construct a training set. A denoising network construction module is used to construct a denoising network. The denoising network is based on the 3D-UNet network architecture and includes an encoder, a bottleneck layer, and a decoder. A multi-scale feature fusion module and a residual block are added sequentially between the input layer and downsampling module of the original 3D-UNet encoder, and an edge enhancement module is integrated into the residual block. Specifically, the multi-scale feature fusion module extracts multi-scale spatial-spectral features from the input data; the edge enhancement module enhances the edges of the multi-scale spatial-spectral features; the downsampling module uses a parallel-branch convolutional structure to extract the downsampling features of the enhanced features and the spatial dimension detail features, which are then fused to obtain the fused features; the bottleneck layer extracts the global semantic features of the fused features, which are then processed by the decoder to output predicted noise. The denoising network training module is used to take clean background pixels with randomly added Gaussian noise as noisy samples and input them into the denoising network to obtain predicted noise; construct a denoising loss based on the difference between the predicted noise and the Gaussian noise; reconstruct the background image based on the predicted noise estimate and construct a reconstruction loss based on the difference between the reconstructed background image and the clean background pixels; and iteratively optimize the denoising network through the denoising loss and the reconstruction loss to obtain the trained denoising network. The deep space debris anomaly detection module is used to input the deep space exploration image to be detected into the trained denoising network and output predicted noise; based on the predicted noise and the deep space exploration image to be detected, it estimates and reconstructs a clean background image; it performs Mahalanobis distance detection on the reconstructed clean background image to determine the anomaly score of each pixel in the reconstructed clean background image; when the anomaly score is higher than a preset threshold, the pixel is determined to be a space debris.

Images